Loading…
Protein inverse folding for low-resource enzyme design, distilling a frozen protein language model into a structure encoder used alone at inference.
No providers recorded yet. Browse all providers
Inverse folding models learn from paired structures and sequences, and those pairs are scarce: the standard CATH 4.2 split offers 18,204 training examples, while the sequence databases that protein language models pretrain on hold orders of magnitude more. A structure encoder is starved in exactly the regime where a sequence model is rich, and the gap widens for functional classes such as enzymes, where experimentally determined structures are rarer still.
CrossDesign closes that gap by treating computational protein design as a cross-modal alignment problem. During training it runs two streams in parallel: a structure-to-sequence stream that reads a backbone and writes an amino-acid sequence, and an auxiliary stream in which a frozen, off-the-shelf pretrained protein language model reads the native sequence. The two streams share a decoder, and knowledge-distillation losses pull the structure stream's internal representations toward the language model's. Once training finishes the language model is discarded entirely — inference runs the structure stream alone, so the pretrained prior costs nothing at sampling time.
The model was developed by Jiangbin Zheng, Ge Wang, Han Zhang and Stan Z. Li at Westlake University and Zhejiang University, and presented at ICASSP 2025. It targets low-resource functional design: the authors curated two enzyme benchmarks alongside the generic protein evaluations, and applied a single trained checkpoint to both without enzyme-specific retraining.
The structure-to-sequence stream begins with a transformation-enhanced geometric vector perceptron (tGVP) encoding rotation- and translation-invariant local frame features from backbone N, Cα and C coordinates. A Temporal GraphTransformer — a graph convolution network followed by a Transformer encoder — links that geometric module to a Transformer decoder with learnable positional embeddings. The auxiliary stream passes the native sequence through the frozen language model and a two-hidden-layer MLP adapter into the same decoder, trained as an asymmetric autoencoder. Primary generation uses an exponential cross-entropy loss that sharpens the output distribution.
Trained only on the 18,204-pair CATH 4.2 split, CrossDesign reaches 3.67 perplexity and 57.11% amino-acid recovery on the full CATH test set, ahead of ProteinMPNN at 45.96% and PiFold at 51.66%. It also exceeds ESM-IF (52.51%), which trained on CATH plus predicted AlphaFold DB structures — the paper's central evidence that transferred sequence priors can substitute for structural training scale. On the held-out Ts50 and Ts500 sets recovery reaches 60.04% and 60.90%. Enzyme recovery holds up out of domain: 63.86% on the PET evaluation set and roughly 61% across EnzFoldDB, ranging from 59.52% to 63.18% by fold. On all 217 ProteinGym deep mutational scanning proteins, non-autoregressive scoring attains a mean Spearman ρ of 0.445, against 0.437 for the strongest language model baseline (VESPA) and 0.422 for ESM-IF, with its largest margins on stability (0.636) and binding (0.412) assays.
The primary use is redesigning enzyme sequences on a fixed backbone — diversifying a PET hydrolase or another fold-specified biocatalyst without a family-specific model. Because the same checkpoint also scores mutants, an engineering campaign can use one model both to propose full-sequence redesigns and to rank point or multi-point mutants of a characterized wild type, narrowing a library before synthesis. The fold-partitioned EnzFoldDB results are useful to groups working on one specific fold, reporting recovery per fold rather than a single pooled number.
CrossDesign's contribution is methodological: a sequence-only pretrained model can serve as a training-time teacher for a structure encoder and then be dropped, converting a large pretrained prior into a permanent gain in a small, deployable model. The enzyme benchmarks are a useful artifact in their own right, since function-specific structure-sequence evaluation sets remain rare. The limits are real. Neither code nor weights have been released and no license is stated, so the results are not independently reproducible and the model cannot be run by others. The authors note that the functional protein datasets are small, and that the mutation-effect predictions rest on in-silico correlation with deep mutational scanning data, without wet-lab validation of any designed or scored sequence.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.